Comparison of macular GCIPL and peripapillary RNFL deviation maps for detection of glaucomatous eye with localized RNFL defect

Comparison of macular GCIPL and peripapillary RNFL deviation maps for detection of glaucomatous eye with localized RNFL defect Mi Jeung Kim, 1,2 Ki Ho Park, 1,2 Beong Wook Yoo, 3 Jin Wook Jeoung, 1,2 Hee Chan Kim 4 and Dong Myung Kim 1,2 1 Department of Ophthalmology, Seoul National University College of Medicine, Seoul, Korea 2 Department of Ophthalmology, Seoul National University Hospital, Seoul, Korea 3 Interdisciplinary Program, Bioengineering Major, Graduate School, Seoul National University, Seoul, Korea 4 Department of Biomedical Engineering, College of Medicine and Institute of Medical & Biological Engineering, Medical Research Center, Seoul National University, Seoul, Korea ABSTRACT. Purpose: To evaluate the ability of the deviation map of macular ganglion cell inner plexiform layer (GCIPL) thickness compared with that of peripapillary retinal nerve fibre layer (prnfl) thickness for detection of localized RNFL defects shown on red-free RNFL photography. Methods: This prospective cross-sectional study included 69 eyes of 69 subjects with preperimetric or perimetric glaucoma (mean deviation (MD) > 12dB) and localized RNFL defects along with 79 eyes of 79 normal control subjects. The number of abnormal superpixels on the both macular GCIPL and prnfl deviation maps by Cirrus OCT corresponding to localized RNFL defects was calculated using a customized Matlab program and presented as severity indices according to each of the probability levels. The areas under the receiver operating characteristic curves (AUROCs) of the severity indices were compared between the two deviation maps. Results: According to three criteria and corresponding probability levels, the AUROCs of the GCIPL and prnfl deviation maps ranged from 0.910 to 0.931 and 0.934 to 0.950, respectively. However, the differences were not statistically significant (all p > 0.05). Significant correlations were observed between the severity indices of the GCIPL deviation map and those of the prnfl deviation map, regardless of the criteria (all p < 0.0001). Conclusions: In the detection of glaucomatous eyes with localized RNFL defects, the macular GCIPL thickness deviation map showed a level of diagnostic performance comparable to that of the prnfl thickness deviation map. Key words: deviation map ganglion cell analysis ganglion cell inner plexiform layer retinal nerve fibre layer Acta Ophthalmol. 2015: 93: e22 e28 ª 2014 Acta Ophthalmologica Scandinavica Foundation. Published by John Wiley & Sons Ltd doi: 10.1111/aos.12485 Introduction is a progressive optic neuropathy characterized by the loss of retinal ganglion cells (RGCs) presenting as structural changes of the optic nerve head (ONH) and retinal nerve fibre layer (RNFL) with corresponding visual field (VF) defects (Takayama et al. 2012). For prevention of functional vision loss and maintenance of the quality of life of glaucoma patients, it is essential to detect early change by glaucomatous RGC loss. Previous study has reported that peripapillary RNFL (prnfl) thickness measurement by Cirrus highdefinition (HD) spectral-domain (SD) Optical Coherence Tomography (OCT) (software version 5.1.0.96; Carl Zeiss Meditec, Dublin, CA, USA) is highly effective for detection of localized RNFL defect on red-free RNFL photography (Hwang et al. 2013). The prnfl deviation map (Cirrus OCT software version 3.0; Carl Zeiss Meditec) has been shown to offer a higher diagnostic ability for detection of localized RNFL defects than other conventional prnfl maps such as the clock-hour map or quadrant map (Jeoung & Park 2010; Kim et al. 2010; Leung et al. 2010). Recently the ganglion cell analysis (GCA) algorithm was incorporated into Cirrus OCT with the newer software version 6.0 to allow for successful and reproducible segmentation of inner macular layers (GCIPL: a combination of the ganglion cell layer [GCL] and the inner plexiform layer [IPL]) (Mwanza et al. 2011a,b); its glaucoma diagnostic ability, significantly, has been shown to be comparable to those of the prnfl and ONH parameters e22

(Kotowski et al. 2012; Mwanza et al. 2012). However, to date, little of the ability of the macular GCIPL deviation map in Cirrus OCT for detection of localized RNFL defects has been demonstrably determined. Therefore, this study was undertaken to compare the ability of the macular GCIPL deviation map with that of the prnfl deviation map for detection of localized RNFL defects shown on red-free RNFL photography. Material and Methods Study subjects tous eyes with localized RNFL defects and normal control eyes satisfying the eligibility criteria were enrolled from the Clinic of Seoul National University Hospital during the period from December 2012 to May 2013. In cases where both eyes were eligible, one eye was randomly selected. The subjects were all aged 18 years or older. All had a best-corrected visual acuity (BCVA) of 20/40 or better, a spherical-equivalent refractive error within 5.00 dioptres (D), astigmatism within 3.00 D, an open anterior chamber angle and high-quality redfree RNFL photography. Eyes with a history of ocular or systemic diseases possibly affecting the peripapillary area (e.g. large peripapillary atrophy, chorioretinal coloboma, peripapillary staphyloma) or macula area (e.g. epiretinal membrane, age-related macular degeneration, macular oedema), amblyopia, uveitis, intra-ocular surgery (excepting uncomplicated cataract surgery), diabetes or any other ocular or systemic diseases affecting RNFL thickness or VF (e.g. retinal vein occlusion, ischemic optic neuropathy) were excluded. OCT images with a signal strength 6 or with visible eye motion or blinking artifacts were excluded as well. The glaucomatous eyes had localized RNFL defects on red-free RNFL photography with asymptomatic-to-moderate glaucomatous VF loss (mean deviation (MD) > 12 db). tous VF loss was defined as a pattern standard deviation (PSD) outside the 95% normal limits, glaucoma hemifield test results outside the normal limits and/or a cluster of at least three points with a p-value <0.05 on the pattern deviation plot, 1 of each with p < 0.01 affecting the same hemifield; also, the cluster could not be contiguous with the blind spot and could not cross the horizontal midline, on two consecutive VF tests. Preperimetric glaucomatous eyes were defined as those having a localized wedge-shaped RNFL defect clearly visible on red-free RNFL photography with normal standard automated perimetry (SAP) results in at least two tests. Normal VF was defined as MD and PSD within 95% confidence limits and a glaucoma hemifield test result within the normal limits. Normal control eyes were defined as those having an intra-ocular pressure (IOP) 21 mmhg with no history of increased IOP, showing an absence of glaucomatous disc appearance, no visible RNFL defect on red-free RNFL photography and a normal SAP result. The study adhered to the tenets of the Declaration of Helsinki and was approved by the Institutional Review Board of Seoul National University Hospital. Informed consent was obtained from all of the subjects. Red-free RNFL photography The subjects underwent a comprehensive ophthalmic examination including a medical history review; measurement of BCVA; slit-lamp biomicroscopy; Goldmann applanation tonometry (GAT); gonioscopy; dilated fundoscopic examination with a 90 (D) lens; and stereoscopic disc photography (SDP), red-free RNFL photography and SAP (Humphrey Field Analyzer II; Carl Zeiss Meditec). Red-free RNFL photography was obtained after dilation of the pupil using a digital fundus camera system (CF-60UVi/D60; Canon, Inc, Tokyo, Japan) with a green filter inserted to enhance the RNFL (Airaksinen & Nieminen 1985). Images were saved in a 1600 9 1216-pixel digital imaging and communications in medicine format and were stored in the picture archiving communication system (PACS) of Seoul National University Hospital. Localized RNFL defects on red-free RNFL photography were defined as having a width at a 1-disc-diameter distance from the edge of the disc larger than that of a major retinal vessel, diverging in an arcuate or wedge shape and reaching the edge of the disc (Hoyt et al. 1973). Two masked glaucoma specialists (MJK and KHP) independently evaluated the red-free RNFL photography without knowledge of clinical information such as OCT or VF test results; discrepancies were resolved by adjudication of a third glaucoma specialist (JWJ). Cirrus OCT imaging Using Cirrus OCT (software version 6.0), OCT images were acquired by macular scan (macular cube 200 9 200 protocol) and prnfl scan (optic disc cube 200 9 200 protocol) subsequent to pupil dilation. The macular GCIPL thickness within a 6 9 6 9 2mm (14.13 mm 2 ) elliptical annulus around the fovea was measured and computed by GCA algorithm embedded in Cirrus OCT software version 6.0. The annulus cube was of 1 mm inner vertical diameter, 4 mm outer vertical diameter, 1.2 mm inner horizontal diameter and 4.8 mm outer horizontal diameter, excluding the central portions of the fovea where the layers are thin and difficult to defect (Mwanza et al. 2011a). GCIPL thickness was then analysed according to eight parameters: average, minimum and in six sectors (superonasal, superior, superotemporal, inferotemporal, inferior, and inferonasal). This computation method has been described in detail in previous reports (Mwanza et al. 2011a,b). The prnfl thickness within a 3.46-mm diameter circle (256 A-scan) automatically positioned around the optic disc was measured and analysed in 17 parameters: average, four quadrants (superior, inferior, temporal, nasal) and 12 clock-hour sectors (Savini et al. 2011). The RNFL thickness in a 6 9 6mm 2 area around the optic disc cube was measured by 200 9 200 axial scans (pixels) for generation of the deviation map. On the basis of a comparison with the built-in internal normative database, the GCIPL and prnfl thickness values were analysed and then represented on colour-coded deviation maps composed of 50 9 50 superpixels (200 9 200 pixels). The uncoloured (grey colour) superpixels indicated the normal range, whereas yellow- or red-coloured superpixels indicated abnormality at the 5% or 1% level, respectively. The GCIPL deviation map represented the OCT enface image of the annular cube (between the inner and outer rings) of e23

the macula excluding the central fovea. The prnfl deviation map represented the OCT enface image of the optic disc that showed the boundaries of the cup, disc and 3.46-mm diameter circle. Deviation map analysis protocol In this study, the GCIPL and prnfl measurement data were exported as image files (file format: JPEG) using the built-in export function of Cirrus OCT. The deviation map for each file was retrieved using a customized image processing program written in Matlab R2012a (The Mathworks, Inc., Natick, MA, USA). Then, the number of colour-coded (yellow or red) abnormal superpixels on the GCIPL and prnfl deviation map corresponding to the location of localized RNFL defects visible on the red-free RNFL photographs was calculated using the same customized Matlab program. The three criteria for significantly aberrant superpixels on the GCIPL and prnfl deviation map corresponding to RNFL defect were determined arbitrarily as follows: Criterion 1: Cluster of 3 contiguous yellow superpixels including 1 red superpixel. Criterion 2: Cluster of 5 contiguous yellow superpixels including 3 red superpixels. Criterion 3: Cluster of 10 contiguous yellow superpixels including 5 red superpixels. In the case of the macular GCIPL deviation map, we excluded from the analysis 1 superpixel around the inner circle of the GCIPL scan area considering the possibility of artifact. With respect to the prnfl deviation map, we excluded the nasal quadrant (90 ) centred on the prnfl calculation circle (3.46-mm diameter) as well as three superpixels around the disc margin, in the light of the frequency of glaucomatous damage and the possibility of artifact (Fig. 1). The number of aberrant superpixels was calculated according to the severity index (S1, S2, or S3), which was determined in relation to the probability level. S1 represented the number of yellow and red superpixels for the p < 0.05 probability level; S2 indicated the number of only red superpixels for the p < 0.01 probability level. As the red superpixels represented a more significant abnormality than the yellow superpixels, we double-weighted the red superpixels. Therefore, S3 represented the number of red superpixels multiplied by 2 plus the number of yellow superpixels. Statistical analysis Statistical analyses were performed using SPSS version 19.0 (SPSS Inc., Chicago, IL, USA) and MedCalc 12.3.0 (MedCalc Software, Mariakerke, Belgium). The Student t-test was used to compare the continuous variables between the normal control group and the glaucoma group. Additionally, the Pearson chi-square test was used to compare the two groups categorical variables. Pearson correlation analyses were performed to calculate the correlations of the deviation map results (the number of aberrant superpixels represented as a severity index for each probability level: S1, S2, S3) between the GCIPL and prnfl. The diagnostic abilities of the deviation map algorithms in discriminating glaucomatous eyes with localized RNFL defects from normal control eyes were evaluated by computing the areas under the receiver operating characteristic curves (AU- ROCs) and comparing the results (this procedure was described by DeLong et al. 1988). A p-value <0.05 was considered statistically significant. Results The final study sample included 148 eyes of 148 subjects (69 eyes of 69 subjects with localized RNFL defects and 79 eyes of 79 normal control subjects). Among those 69 eyes with localized RNFL defects, 51 showed perimetric glaucoma (40 eyes: early stage (MD > 6 db) of VF loss; 11 eyes: moderate stage ( 12 db< MD 6 db) of VF loss) and 18 showed preperimetric glaucoma. The ocular and demographic characteristics of the subjects are presented in Table 1. Between the two groups, there were no significant differences in mean age, BCVA or spherical equivalent of refraction (all p > 0.05, Student t-test). However, the sex distribution showed a significant discrepancy, in that more females were included in the glaucoma group (64%) than in the normal control group (47%) (p = 0.039, chisquare test). Moreover, the VF indices (MD, PSD) did significantly differ (p < 0.0001, Student t-test). Table 2 shows the averages of the severity indices for the GCIPL and prnfl deviation maps as determined by the number of aberrant superpixels at each probability level. Comparing the glaucoma groups and the normal control group, there were statistically significant differences in all three severity indices (S1, S2, S3) for each criterion (all p < 0.0001, Student t-test). The severity indices of both deviation maps generally decreased according to the change of criterion from 1 to 3. However, the S2 (number of red superpixels) of both maps showed relatively similar values regardless of the criterion. With the GCIPL deviation map algorithm, the AUROCs in discriminating the glaucomatous eyes with localized RNFL defects from the normal control eyes ranged from 0.910 to 0.931 and for the prnfl deviation map algorithm, from 0.934 to 0.950, according to the criteria and probability levels. In a comparison of all of the corresponding severity indices, those for the prnfl deviation map showed larger AUROCs than those for the GCIPL deviation map. However, the differences were not statistically significant (all p > 0.05; Fig. 2). According to the different criteria, the AUROCs for S2 (0.931 for Criterion 1, 0.924 for Criterion 2 and 0.916 for Criterion 3) and S3 (0.922 for Criterion 1, 0.917 for Criterion 2 and 0.910 for Criterion 3) of the GCIPL deviation map were significantly larger than those for S1 (all p < 0.05); there was no significant difference between S2 and S3 (all p > 0.05). By contrast, there were no significant differences found among the three severity indices of the prnfl deviation map for any of the criteria (all p > 0.05). When matched for fixed specificity, the sensitivities of the severity indices of the prnfl deviation map were generally higher than those of the GCIPL deviation map, excepting S2 for criteria 2 and 3 (specificity 95%; Table 3). To investigate the correlation of the severity indices in the deviation map algorithm between the GCIPL and prnfl, Pearson correlation analysis was performed. Strongly significant correlations were observed in the three indices, regardless of the criterion (all p < 0.0001). S3 showed the strongest correlations between the GCIPL deviation map and the prnfl deviation e24

(A) (B) (C) Fig. 1. Calculation of aberrant deviation-map superpixels corresponding to localized retinal nerve fibre layer (RNFL) defect using customized Matlab program. (A) Red-free RNFL photography showing inferotemporal localized RNFL defect (arrowheads). (B) Aberrant prnfl deviationmap superpixels corresponding to photographic RNFL defect: nasal 90 centred on the prnfl calculation circle and three superpixels near the disc margin were excluded from the analysis (blue line). The superpixels coded in red and yellow were recognized and calculated by the customized image processing program (red superpixels for the area bounded by the green line, yellow superpixels for the area bounded by the orange line). (C) Aberrant GCIPL deviation-map superpixels corresponding to photographic RNFL defect: 1 superpixel near the inner ring (border) of the GCIPL scan area (purple line) was excluded from the analysis (blue line). The superpixels coded in red and yellow were recognized and calculated by the customized image processing program (red superpixels for the area bounded by the green line, yellow superpixels for the area bounded by the orange line). Table 1. Demographics and ocular characteristics of normal controls and subjects with glaucoma and localized retinal nerve fibre layer (RNFL) defects. map (coefficient of determination (R 2 ) = 0.642 for Criterion 1, R 2 = 0.653 for Criterion 2 and R 2 = 0.654 for Criterion 3; all p < 0.0001) compared with S2 (R 2 = 0.627 for Criterion 1, R 2 = 0.629 for Criterion 2 and R 2 = 0.626 for Criterion 3; all p < 0.0001) and S1 (R 2 = 0.615 for Criterion 1, R 2 = 0.632 for Criterion 2 and R 2 = 0.638 for Criterion 3; all p < 0.0001; Table 4). Discussion The present study was designed with the main objective of evaluating the diagnostic ability of the GCIPL deviation map in Cirrus OCT for detection of localized RNFL defects. The results Normal control (n = 79) (n = 69) p-value Age (y) (mean SD) 56.14 12.27 58.19 9.91 0.270* Sex (male/female) 42/37 25/44 0.039 BCVA (log MAR) 0.04 0.65 0.04 0.74 0.802* SE (dioptres) 0.73 1.51 0.27 1.62 0.143* VF Index MD (db) 0.44 1.90-3.02 3.21 <0.0001* PSD (db) 2.02 1.08 5.35 3.88 <0.0001* Stage of disease Preperimetric 18 (26%) Early (MD > 6 db) 40 (58%) Moderate ( 12 db < MD 6 db) 11 (16%) SD = standard deviation, BCVA = best-corrected visual acuity, SE = spherical equivalent, VF = visual field, MD = mean deviation, PSD = pattern standard deviation. * The comparison was performed using the Student t-test. The comparison was performed using the chi-square test. showed that the GCIPL deviation map offers favourable, comparable-toprnfl diagnostic performance in the detection of such defects on red-free RNFL photography. In a large proportion of glaucoma patients, RNFL damage is known to precede noticeable change of the ONH and VF defect (Sommer et al. 1977, 1991; Tuulonen et al. 1993). Therefore, RNFL assessment is quite helpful for early diagnosis of glaucoma. There are various techniques available for detecting RNFL defects; they include fundus examination, red-free RNFL photography, Heidelberg Retina Tomography (HRT) and OCT. Given the recent advances in OCT technology, SD-OCT offers objective, reproducible real-time measurement of RNFL thickness with a fast scan speed (Wojtkowski et al. 2005; Leung et al. 2009; Zhao et al. 2014). Notably, Cirrus OCT, a commercial SD-OCT platform, generates a prnfl deviation map that outperforms the conventional clock-hour map or quadrant map in RNFL defect detection (Jeoung & Park 2010; Kim et al. 2010; Leung et al. 2010). For example, Hwang et al. (2013) reported that for 295 eyes with early-stage glaucoma (MD > 6.0 db), the prnfl deviation map by Cirrus OCT showed a lower frequency of misidentification of photographic RNFL defects than the clock-hour map. Meanwhile, the GCA algorithm, recently incorporated into Cirrus OCT, provides the macular GCIPL deviation map by analysing the GCIPL thickness for each superpixel in an area around the fovea (Mwanza et al. 2011a,b). Although there have been a few studies on the diagnostic ability of the GCIPL deviation map for discriminating glaucomatous eyes from normal control eyes, they have shown only whether the GCIPL deviation map detects glaucomatous damage or not (Kotowski et al. 2012; Mwanza et al. 2012; Sung et al. 2013). For example, Sung et al. (2013) have reported that the GCIPL deviation map showed similar discrimination ability between normal controls and patients with preperimetric or early glaucoma and compared with prnfl deviation map by quantifying abnormal superpixels. However, to our knowl- e25

Table 2. Severity index of deviation map of macular ganglion cell inner plexiform layer (GCIPL) and peripapillary retinal nerve fibre layer (prnfl) computed using a customized Matlab program. S1 (red and yellow superpixels) S2 (red superpixels only) S3 (red superpixels 9 2 + yellow superpixels) Criteria Normal control (n = 79) (n = 69) p- value* Normal control (n = 79) (n = 69) p- value* Normal control (n = 79) (n = 69) p- value* Criterion 1 prnfl 16.01 37.77 196.93 111.12 <0.0001 4.39 13.11 111.30 71.17 <0.0001 20.41 47.99 308.23 179.35 <0.0001 GCIPL 34.66 69.50 275.39 163.52 <0.0001 7.76 17.61 204.01 151.27 <0.0001 42.42 85.14 479.41 311.80 <0.0001 Criterion 2 prnfl 10.29 29.91 192.96 111.13 <0.0001 4.04 13.10 111.70 72.38 <0.0001 14.33 41.87 304.65 180.17 <0.0001 GCIPL 30.22 66.12 271.61 166.48 <0.0001 7.13 16.88 203.28 151.28 <0.0001 37.34 81.03 474.88 314.79 <0.0001 Criterion 3 prnfl 8.10 27.35 191.57 110.79 <0.0001 3.57 12.85 110.03 71.30 <0.0001 11.67 39.30 301.59 179.20 <0.0001 GCIPL 26.14 61.91 270.28 167.39 <0.0001 6.24 16.32 202.68 151.56 <0.0001 32.38 76.32 472.96 315.91 <0.0001 S1 = number of yellow and red superpixels with probability level of p < 0.05, S2 = number of only red superpixels with probability level of p < 0.01, S3 = number of red superpixels 9 2 + number of yellow superpixels, Criterion 1 = Cluster of 3 contiguous yellow superpixels including 1 red superpixel; Criterion 2 = Cluster of 5 contiguous yellow superpixels including 3 red superpixels; Criterion 3 = Cluster of 10 contiguous yellow superpixels including 5 red superpixels. * Comparison was performed using Student t-test. (A) (B) (C) Fig. 2. AUROC of severity indices on deviation map for macular ganglion cell inner plexiform layer (GCIPL) and peripapillary retinal nerve fibre layer (prnfl) by defined criteria. (A), AUROC of S1 on both GCIPL and prnfl deviation maps. (B), AUROC of S2 on both deviation maps. (C), AUROC of S3 on both deviation maps. edge, there has been no study evaluating the ability of GCIPL deviation map to identify localized RNFL defect and comparing with that of prnfl deviation map. In the present study, for the purposes of a quantitative analysis on the ability to detect localized RNFL defect, we digitized the deviation map results so that the number of abnormal superpixels could be counted by a customized image processing program. Moreover, we used completely automated image processing program for counting the number of abnormal superpixels, which entailed the advantage of eliminating the counting errors and interobserver variability that would be incurred when counting manually or using software requiring manual determination of the boundaries of deviated superpixels. In our results, all of the severity indices (S1, S2, and S3), which represented the number of abnormal superpixels counted by the image processing program, were larger in glaucomatous eyes with RNFL defects than in normal control eyes (Table 1). This means that the results of the deviation map corresponded well to the RNFL defects visible on red-free RNFL photography. As the scan area of the GCIPL deviation map was limited to the parafoveal region and included only about 50% of RGC population (Curcio & Allen 1990), it might not detect RNFL defects located far from the fovea. Therefore, we assumed that the ability of the GCIPL deviation map for detection of localized RNFL defects might prove inferior to that of the prnfl deviation map that computes data on the full 360-degree peripapillary region, including a total sampling of the RGC axons. We found that the AUROCs for detection of localized RNFL defects were larger on the prnfl deviation map than on the GCIPL deviation map. However, the differences were not statistically significant. This was contrary to our earlier expectation (Table 3). Determining the exact mechanism involved will require further investigation stratified by the peripapillary widths and locations of RNFL defects. We found that S2, which indicated the number of red superpixels (abnor- e26

Table 3. Areas under receiver operating characteristic curves (AUROCs) and sensitivities at fixed specificities for severity indices on deviation map of macular ganglion cell inner plexiform layer (GCIPL) and peripapillary retinal nerve fibre layer (prnfl). Sensitivity AUROC SE (95% CI) Specificity 80% Specificity 95% GCIPL prnfl p-value* GCIPL prnfl GCIPL prnfl Criterion 1 S1 0.915 0.024 (0.868 0.961) 0.934 0.023 (0.889 0.978) 0.468 87.0 91.3 62.3 75.4 S2 0.931 0.022 (0.888 0.974) 0.943 0.022 (0.900 0.986) 0.633 85.5 92.8 79.7 81.2 S3 0.922 0.023 (0.878 0.967) 0.938 0.022 (0.895 0.982) 0.526 87.0 92.8 72.5 79.7 Criterion 2 S1 0.910 0.025 (0.860 0.959) 0.944 0.021 (0.903 0.985) 0.133 88.4 92.8 68.1 79.7 S2 0.924 0.024 (0.877 0.971) 0.946 0.021 (0.906 0.987) 0.310 85.5 92.8 81.2 79.7 S3 0.917 0.025 (0.868 0.965) 0.946 0.021 (0.906 0.987) 0.184 88.4 92.8 72.5 82.6 Criterion 3 S1 0.904 0.027 (0.851 0.957) 0.949 0.020 (0.910 0.989) 0.063 88.4 92.8 68.1 87.0 S2 0.916 0.026 (0.865 0.967) 0.949 0.020 (0.910 0.989) 0.159 85.5 92.8 81.2 78.3 S3 0.910 0.026 (0.858 0.962) 0.950 0.020 (0.910 0.989) 0.093 88.4 92.8 72.5 82.6 AUROC = area under receiver operating characteristic curve, SE = spherical equivalent, S1 = number of yellow and red superpixels with probability level of p < 0.05, S2 = number of only red superpixels with probability level of p < 0.01, S3 = number of red superpixels 9 2 + number of yellow superpixels, Criterion 1 = Cluster of 3 contiguous yellow superpixels including 1 red superpixel, Criterion 2 = Cluster of 5 contiguous yellow superpixels including 3 red superpixels, Criterion 3 = Cluster of 10 contiguous yellow superpixels including 5 red superpixels. * Comparison was performed by the method of DeLong et al. (1988). Table 4. Correlation of severity index of deviation map algorithm between macular ganglion cell inner plexiform layer (GCIPL) and peripapillary retinal nerve fibre layer (prnfl) by Pearson correlation analysis. Criteria GCIPL S1 versus prnfl S1 mal at the probability level <0.01) and S3, which weighted the number of red superpixels, showed significantly larger AUROCs than S1. Therefore, we suggest observing red superpixels more closely than yellow ones on the deviation map would be better for GCIPL deviation map evaluation. In addition, the values of sensitivities at fixed specificities (specificity > 95%) in this study generally were not high, which was ranged from 62.3% to 81.2% on the GCIPL deviation map and from 75.4% to 87.0% on the prnfl deviation map. We speculated that this was due to the fact that we included patients with mostly early stage of glaucoma (40 eyes; 58%), even with preperimetric stage of glaucoma (18 eyes; 26%). Because the diagnostic performance of imaging device is highly affected by the severity of disease (Leite et al. 2010), the relatively low diagnostic performance may be GCIPL S2 versus prnfl S2 GCIPL S3 versus prnfl S3 R 2 p-value R 2 p-value R 2 p-value Criterion 1 0.615 <0.0001 0.627 <0.0001 0.642 <0.0001 Criterion 2 0.632 <0.0001 0.629 <0.0001 0.653 <0.0001 Criterion 3 0.638 <0.0001 0.626 <0.0001 0.654 <0.0001 R 2 : Coefficient of determination by Pearson correlation analysis. related to the early stage of glaucomatous damage in study subjects. All of the three severity indices were found to be significantly and closely correlated between the two deviation maps (GCIPL and prnfl) (Table 4). That is, the more abnormal superpixels on the GCIPL deviation map corresponded to RNFL defects, the more abnormal superpixels on the prnfl deviation map also corresponded to RNFL defects. Considering the pathophysiology of the glaucomatous optic neuropathy that primarily affected the RGC and their axons (RNFL), we cautiously speculated that this result was due to the fact that the (glaucomatous) RNFL defect, presenting as abnormal superpixels on the prnfl deviation map, was mostly accompanied by the RGC loss presenting as abnormal superpixels on the GCIPL deviation map. In this context, we supposed that abnormal lesions corresponding to RNFL defect on the two deviation maps would tend to coincide. In other words, glaucomatous structural damage in these two anatomical lesions would be detected by both GCIPL and prnfl deviation maps, respectively. In this study, the assessment of diagnostic ability was performed under a case control design, which included eyes with a photographically identifiable (noticeable) RNFL defect (the cases) and normal eyes with a perfectly normal RNFL (the controls). Because the RNFL defects met certain levels of depth and width to qualify as defects on red-free RNFL photography, distinctly glaucomatous eyes with prominent structural changes were enrolled in the present study. Thus, in this study, the diagnostic performance of the deviation map for both scans might have been overestimated relative to real clinical practice, in which veryearly-stage glaucomatous eyes with ambiguous RNFL defects are encountered. However, in present study, among 69 eyes with localized RNFL defects, 18 (26.1%) showed asymptomatic VF loss (preperimetric glaucoma) and 40 (58.0%) early-stage VF loss (MD > 6 db). Considering that 58 of the 69 eyes (84.1%) showed early-stage VF loss, we supposed that the influence of subject selection on the study results was not significant. The present study has several limitations. First, as aforementioned, we e27

integrated all of the subjects within a preperimetric-to-moderate range of glaucoma. Because the diagnostic ability was influenced by the disease severity, for further, high-precision investigation of the diagnostic accuracies of those deviation maps, subgroup analysis according to disease severity, with a large number of cases, will be mandatory. Second, we excluded diffuse RNFL atrophies and ambiguous RNFL defects from consideration, enrolling only patients with localized RNFL defects and clear margins. Considering, however, that glaucoma patients presenting with solely localized RNFL defects represent only a subset of the total glaucoma patient population, for more comprehensive evaluation of the utility of the GCIPL deviation map for detection of RNFL defects in the clinical field, further study involving defects without clear margins is required. Third, we excluded eyes with a history of ocular or systemic disease possibly affecting the peripapillary area or macula area, and all of our OCT images were obtained after dilation of the pupil. However, in the real-world clinical setting, OCT is frequently performed on eyes with several minor lesions in the macula or peripapillary area without pupil dilation. Therefore, the findings of this study, in the context of real clinical practice, might be limited. Fourth, the criteria for significantly aberrant superpixels corresponding to RNFL defects used in this study were arbitrary, lacking any anatomical explanation for the cut-off number of superpixels. To compensate for this weakness, we established not a single criterion but rather three criteria for the diversity and the possibility. Consequently, all the analysis showed consistent results, regardless of the criterion. These limitations notwithstanding, the present study remains significant for its first quantitative comparison of the diagnostic abilities of the GCIPL and prnfl deviation maps in detecting localized RNFL defects. A quantitative analysis of the respective deviation map algorithms using a customized image processing program showed comparable diagnostic performances for early-to-moderate-stage glaucomatous eyes, though the AU- ROC of the prnfl deviation map showed numerically better values than that of the GCIPL. References Airaksinen PJ & Nieminen H (1985): Retinal nerve fiber layer photography in glaucoma. Ophthalmology 92: 877 879. Curcio CA & Allen KA (1990): Topography of ganglion cells in human retina. J Comp Neurol 300: 5 25. DeLong ER, DeLong DM & Clarke-Pearson DL (1988): Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 44: 837 845. Hoyt WF, Frisen L & Newman NM (1973): Fundoscopy of nerve fiber layer defects in glaucoma. Invest Ophthalmol 12: 814 829. Hwang YH, Kim YY, Kim HK & Sohn YH (2013): Ability of cirrus high-definition spectral-domain optical coherence tomography clock-hour, deviation, and thickness maps in detecting photographic retinal nerve fiber layer abnormalities. Ophthalmology 120: 1380 1387. Jeoung JW & Park KH (2010): Comparison of cirrus OCT and stratus OCT on the ability to detect localized retinal nerve fiber layer defects in preperimetric glaucoma. Invest Ophthalmol Vis Sci 51: 938 945. Kim NR, Lee ES, Seong GJ, Choi EH, Hong S & Kim CY (2010): Spectral-domain optical coherence tomography for detection of localized retinal nerve fiber layer defects in patients with open-angle glaucoma. Arch Ophthalmol 128: 1121 1128. Kotowski J, Folio LS, Wollstein G, Ishikawa H, Ling Y, Bilonick RA, Kagemann L & Schuman JS (2012): discrimination of segmented cirrus spectral domain optical coherence tomography (SD-OCT) macular scans. Br J Ophthalmol 96: 1420 1425. Leite MT, Zangwill LM, Weinreb RN, Rao HL, Alencar LM, Sample PA & Medeiros FA (2010): Effect of disease severity on the performance of Cirrus spectral-domain OCT for glaucoma diagnosis. Invest Ophthalmol Vis Sci 51: 4104 4109. Leung CK, Cheung CY, Weinreb RN et al. (2009): Retinal nerve fiber layer imaging with spectral-domain optical coherence tomography: a variability and diagnostic performance study. Ophthalmology 116: 1257 1263. Leung CK, Lam S, Weinreb RN et al. (2010): Retinal nerve fiber layer imaging with spectral-domain optical coherence tomography: analysis of the retinal nerve fiber layer map for glaucoma detection. Ophthalmology 117: 1684 1691. Mwanza JC, Durbin MK, Budenz DL et al. (2011a): Profile and predictors of normal ganglion cell-inner plexiform layer thickness measured with frequency-domain optical coherence tomography. Invest Ophthalmol Vis Sci 52: 7872 7879. Mwanza JC, Oakley JD, Budenz DL, Chang RT, Knight OJ & Feuer WJ (2011b): Macular ganglion cell-inner plexiform layer: automated detection and thickness reproducibility with spectral domain-optical coherence tomography in glaucoma. Invest Ophthalmol Vis Sci 52: 8323 8329. Mwanza JC, Durbin MK, Budenz DL et al. (2012): diagnostic accuracy of ganglion cell-inner plexiform layer thickness: comparison with nerve fiber layer and optic nerve head. Ophthalmology 119: 1151 1158. Savini G, Carbonelli M & Barboni P (2011): Spectral-domain optical coherence tomography for the diagnosis and follow-up of glaucoma. Curr Opin Ophthalmol 22: 115 123. Sommer A, Miller NR, Pollack I, Maumenee AE & George T (1977): The nerve fiber layer in the diagnosis of glaucoma. Arch Ophthalmol 95: 2149 2156. Sommer A, Katz J, Quigley HA, Miller NR, Robin AL, Richter RC & Witt KA (1991): Clinically detectable nerve fiber atrophy precedes the onset of glaucomatous field loss. Arch Ophthalmol 109: 77 83. Sung MS, Yoon JH & Park SW (2013): Diagnostic validity of macular ganglion cell-inner plexiform layer thickness deviation map algorithm using Cirrus HD-OCT in preperimetric and early glaucoma. J 14: 14. Takayama K, Hangai M, Durbin M, Nakano N, Morooka S, Akagi T, Ikeda HO & Yoshimura N (2012): A novel method to detect local ganglion cell loss in early glaucoma using spectral-domain optical coherence tomography. Invest Ophthalmol Vis Sci 53: 6904 6913. Tuulonen A, Lehtola J & Airaksinen PJ (1993): Nerve fiber layer defects with normal visual fields. Do normal optic disc and normal visual field indicate absence of glaucomatous abnormality?. Ophthalmology 100: 587 597. Wojtkowski M, Srinivasan V, Fujimoto JG, Ko T, Schuman JS, Kowalczyk A & Duker JS (2005): Three-dimensional retinal imaging with high-speed ultrahigh-resolution optical coherence tomography. Ophthalmology 112: 1734 1746. Zhao L, Wang Y, Chen CX, Xu L & Jonas JB (2014): Retinal nerve fibre layer thickness measured by Spectralis spectral-domain optical coherence tomography: The Beijing Eye Study. Acta Ophthalmol 92: 35 41. Received on January 6th, 2014. Accepted on May 24th, 2014. Correspondence: Ki Ho Park, MD, PhD Department of Ophthalmology Seoul National University College of Medicine Seoul National University Hospital 101 Daehak-ro, Jongno-gu Seoul 110-744 Korea Tel: +82 2 2072 2438 Fax: + 82 2 741 3187 Email: kihopark@snu.ac.kr e28